Leading U.S. artificial intelligence labs such as OpenAI and Anthropic are releasing cheaper models, trying to retain customers who are looking to cut costs and are switching to budget alternatives from Chinese competitors. This price war is unfolding against the backdrop of rising AI costs, forcing companies to limit its use and seek cheaper models. This, in turn, helps Chinese developers, including Moonshot and DeepSeek, gain users from Silicon Valley to Europe.
OpenAI recently announced an 80% price cut for GPT-5.6 Luna — its “fastest and most affordable model.” Anthropic launched Claude Opus 5, advertising the system’s “advanced intelligence” at “half the price” of Fable 5, the company’s most powerful model. These moves have contributed to a nearly 25% drop in the prices customers pay for models from leading U.S. labs since mid-July, according to the Silicon Data token price index. Tokens are data units processed by language models and used to calculate many customer bills.
Strategy shift and market impact
The price cuts mark a strategy shift for U.S. AI developers, who create their own “closed” models and have so far competed primarily on performance. The growing power of “open” Chinese models, which developers can freely download and customize, has increased pressure on prices. These changes are also happening as OpenAI and Anthropic plan initial public offerings (IPOs) with trillion-dollar valuations, and investors seek evidence that the industry’s massive AI spending can be profitable.
Corporate AI users face pricing pressure as Anthropic and OpenAI move some corporate clients from fixed subscriptions to pay-per-use, where companies pay according to the computing resources they consume. Some companies have responded to rising bills by setting limits on AI use or testing cheaper alternatives. Companies such as DoorDash and Airbnb have said they have started using Chinese-made models to cut costs.
This transition coincides with a surge in releases from Chinese labs that have narrowed the performance gap with leading U.S. models. This raises concerns in the U.S. tech industry that American developers may lose customers, even while spending significant amounts to maintain their technological edge. AI labs offer a range of models with different capabilities and prices, with costs further depending on the model version and “effort” settings used.
Price change details and comparisons
Customers typically pay for input tokens, used to measure data fed into the model, and output tokens, which measure the volume of information it generates. The latest price cuts from U.S. labs apply to mid-tier products and make them more competitive with Chinese offerings. For example, OpenAI reduced the price of GPT-5.6 Luna from $1 to $0.20 per million input tokens and from $6 to $1.20 per million output tokens.
Anthropic launched Opus 5 at $5 per million input tokens and $25 per million output tokens — half the price of its Fable 5 model. This week, the company canceled a planned price increase for its Sonnet 5 model, which was to take effect in September. However, the stated token prices do not provide a direct comparison between AI models. More powerful models can sometimes complete a task using fewer tokens or with fewer attempts, so a model that appears more expensive at the stated token price may ultimately cost less.
Additionally, most models can operate with different “effort” settings that change the computing power used to respond to a question and can affect both performance and the final cost of completing a task. Artificial Analysis, which evaluates models in areas such as mathematics, science, programming, and reasoning, found that Anthropic‘s Opus 5 at a “medium” effort level demonstrated similar performance and cost per task as Moonshot‘s Kimi K3 at the “maximum” level. OpenAI‘s GPT-5.6 Luna at the “maximum” effort level showed similar performance to DeepSeek‘s V4 Flash at the “maximum” level but cost slightly less than twice as much per task.
Anthropic and OpenAI declined to comment. A person close to Anthropic said the price of Opus 5, lower than the cost of the flagship Fable 5, is part of how the startup’s “model family” is built, so “there is no connection to competitors.” Mantas Lukauskas, head of AI at Hostinger, a web hosting provider that has used large language models since 2020, noted that prices for the best models had been “stable or increasing.” He added that the recent changes in pricing are the “first real test” of whether companies like Anthropic and OpenAI can protect the value of their most advanced offerings: “American labs have cut the middle segment and are protecting the top.”
Source: Ars Technica
